# Independent review of the extension

Owner: validation agent. Updated 2026-11-11. The physical holdout below has been released; it cannot be reused as a fresh unseen test after this point. Other reserved papers are source-selection research copies, not an evaluation suite.

Subsequent fault refinements, including the failed segmented-rotor field criteria and the completed conducting-ring comparison, are reviewed in [FAULT_REFINEMENT_REVIEW.md](FAULT_REFINEMENT_REVIEW.md). That supplement preserves the later dispositions without relabeling the physical holdout.

## Physical current-to-field holdout

A limited physical comparison was completed on the documented SmCo DC fixture of Sjokvist and Eriksson (2014). Inputs and tolerances were frozen in `builder_inputs_smco.json` and `FIELD_CONTRAST_PROTOCOL.md`; the builder recorded analytical reluctance-circuit predictions before receiving the experimental values. Owner source selection was informed by the publication; this was not a random or double-blind study.

| Opposing DC MMF (A-turn) | Phase-independent DC coil current (A) | Measured field contrast (T) | Predicted contrast (T) | Point error | Frozen assumed sensitivity range (T) |
|---:|---:|---:|---:|---:|---:|
| 2,000 | 4.7281 | −0.14437 ±0.010 | −0.13873 | 3.90% | −0.18253 to −0.09764 |
| 4,000 | 9.4563 | −0.29584 ±0.010 | −0.27746 | 6.21% | −0.36506 to −0.19528 |

The ±0.010 T values describe conservative reading uncertainty of a contrast extracted from the published vector plot. They are not a complete original measurement uncertainty. Both contrasts share the zero-current observation and are correlated.

**Disposition: INCONCLUSIVE physical transfer validation.** Both central values meet the frozen point-error criterion. But the 4,000 A-turn sensitivity interval is 0.16978 T wide, exceeding the prespecified 0.060 T informativeness limit. Coverage by this broad assumed range is weak evidence. No parameters or intervals were retuned after release.

The comparison supports the order and direction of the reversible current-to-gap-field response in one large fixture. It does not validate local magnet H, hot NdFeB, irreversible demagnetization, the motor FE implementation, its spatial 1% criterion, or fault behavior. The precise missing constraints are actual steel differential/RD/TD curves, magnet recoil data, fixture/probe geometry and original measurement uncertainty. The source's fitted Br, coercivity and knee shape were excluded.

Source: [author postprint, printed page 4/Fig. 4](https://uu.diva-portal.org/smash/get/diva2:764071/FULLTEXT01.pdf). The source assessment and distinctions from the other candidate studies are in `BENCHMARK_SOURCE_ASSESSMENT.md`.

Reproduce scoring with `python3 score_field_contrast.py`. It checks the frozen prediction and implementation hashes. `FIELD_CONTRAST_SCORE.json` and `field_contrast_released.csv` contain all quantities and criteria.

## Circuit formulation and numerical verification

Reviewed `../fault/coupled.py` and `../fault/run_fault.py`.

- State current is instantaneous AT = N I. With W mapping two independent currents into a balanced three-phase vector, the circuit equation's W^T W factor and division by N² are dimensionally consistent.
- The rotor pattern moves on a fixed stator-coordinate mesh. The winding operator is fixed, so C^T(A_new−A_old) is the correct winding flux difference. It would need modification for a physically moving winding/mesh operator.
- The Schur-complement Newton signs agree with the coupled field and winding equations. The same winding operator excites the FE source and observes linked flux.
- A separate closed-form forced-RL solution reproduces backward-Euler convergence. Errors normalized by exact peak current are 3.3316%, 1.6847%, 0.84695% and 0.42475% at 48/96/192/384 samples per period. This is verification of the time-integration limit; it does not establish the nonlinear machine's H error.
- Nonlinear field convergence is not time/grid convergence. Fault inception angle, longer horizons where the exposure continues to grow, winding temperature/resistance, and end inductance remain consequential sensitivities. No-load and positive-q motoring starts do not bound all braking or field-weakening events.
- A fixed-speed bolted three-phase short with linear-recoil magnets does not include actual rotor inertia/speed response, switch/lead impedance, eddy-current reaction or irreversible material feedback. It is a declared conditional event, not a worst-case fault certification.

Artifacts: `closed_form_circuit_check.py/.json`, `independent_circuit_check.json`.

## Consequential correction: follow the magnet material through the fault

The first proposed summary took the maximum over time of an instantaneous spatial H quantile. That does not measure the fraction of magnet material that ever crosses a threshold: different locations can be exposed at different times.

The revised `MaterialEnvelope` maps H_parallel to rotor material coordinates, accumulates the maximum opposing field at each material location and then computes the area-weighted 99th percentile. Its four-subtriangle overlay accounts for the alternating triangle diagonal on odd angular shifts.

An independent geometric test identified original FE triangles by barycentric point membership at subtriangle centroids rather than copying the implementation's diagonal labels. Across 25 material quadrilaterals, four subtriangles each, and shifts 0, 1, 5 and 335, maximum mapping error was 2.56e−13 kA/m and the tracked-area error was 1.78e−15 relative. This verifies the indexing/area operation for the tested mesh. It is synthetic numerical verification, not evidence that a knee crossing produces the assumed magnetic loss.

Artifacts: `material_envelope_mapping_check.py/.json`.

## Consequential correction: torque sign

The native Arkkio observer integrates tangential Maxwell stress on an outward cylindrical surface. It reports torque on the interior body. For this outer-rotor machine, that body is the stator, so rotor electromagnetic torque is the negative of the native observer.

At a fixed-current 300 A-turn coarse-grid state, a central difference of the minimized magnetic potential with rotor angle gives +3.05045 N m rotor torque; the native observer gives −3.05577 N m. The approximately 0.17% magnitude difference for ±1 column is consistent with this sign interpretation. The ±2-column difference gives +3.01642 N m and exposes finite-angle dependence. These are internal observer comparisons, not torque validation.

Retain the native observer for reproducibility, but label it and report rotor torque separately for mechanical-energy reasoning. The preceding mass analysis compared torque magnitudes, so this sign correction does not reverse its equal-magnitude roots.

Artifacts: `torque_sign_check.py/.json`.

## Equal-torque mass and axial checks

The mass report's copper accounting is coherent under fixed turns, mean turn length and copper temperature: equal total copper heat requires conductor area and copper mass to scale with squared current ratio. Within that declared model, the thin radial/SUS construction fails the equal-heat material-mass comparison at all three directly solved loads. That conditional negative result does not rule out shorter duty or unused thermal capacity.

The ring-free alternative's 15.93 g combined material credit includes 10.26 g of removed copper. Equal total I²R heat after removing conductor area does **not** preserve local heat transfer, hotspot temperature or transient heat capacity. Its unchanged-winding result—5.68 g rotor material credit and less copper heat—is the cleaner provisional comparison. Both still omit replacement retention and changed losses.

The axial harmonic transfer expression follows the stated scalar-potential boundary-value problem. The axial receiver-window variation is consequential: field outside the active stack need not be lost winding linkage. The model is a legitimate three-coordinate ideal-plane calculation, with independently discretized numerical corroboration. It is not a native slotted finite-iron-end motor solve and must not be used as a multiplicative motor H/torque correction. A 3D correction cannot explain a discrepancy between two 2D FE calculations.

An independent arithmetic rerun from the solved-current CSV reproduced all copper-mass and net-mass values exactly to displayed floating-point precision; largest copper-power difference was 1.14e−13 W. A separately coded change-of-variable spectral integral using the direct hyperbolic transfer expression gives 0.93717580675 for the active-length axial ratio. These checks are recorded in `mass_axial_arithmetic_check.json` and verify the calculations, not their omitted physical inputs.

These reviews find no additional arithmetic defect that reverses the stated conditional findings. The subsequent complete 32-case fault audit, final radial refinement, conducting-ring disposition and bounded investigation closure are recorded in [FAULT_REFINEMENT_REVIEW.md](FAULT_REFINEMENT_REVIEW.md). That later closure preserves the physical holdout's inconclusive result and the failed segmented-rotor refinement criteria.

